1package agent
2
3import (
4 "bufio"
5 "context"
6 "encoding/json"
7 "errors"
8 "fmt"
9 "io"
10 "os"
11 "strings"
12 "sync"
13 "time"
14
15 "github.com/patriceckhart/zot/internal/agent/extensions"
16 "github.com/patriceckhart/zot/internal/agent/modes"
17 "github.com/patriceckhart/zot/internal/core"
18 "github.com/patriceckhart/zot/internal/extproto"
19 "github.com/patriceckhart/zot/internal/provider"
20)
21
22// runRPCMode implements the JSON-over-stdin/stdout RPC protocol.
23//
24// Wire format: one JSON object per line in both directions.
25//
26// Commands (stdin):
27//
28// {"id":"1","type":"prompt","message":"hello","images":[]}
29// {"id":"2","type":"abort"}
30// {"id":"3","type":"compact"}
31// {"id":"4","type":"get_state"}
32// {"id":"5","type":"set_model","model":"claude-opus-4-5"}
33// {"id":"6","type":"get_messages"}
34// {"id":"7","type":"clear"}
35// {"id":"8","type":"get_models"}
36//
37// Responses (stdout): {"type":"response","id":"1","command":"prompt","success":true}
38// Events (stdout): one JSON object per AgentEvent (same schema as --json mode).
39//
40// Auth: if $ZOTCORE_RPC_TOKEN is set, the first command must be
41// {"type":"hello","token":"..."} or the connection is closed.
42func runRPCMode(ctx context.Context, args Args, version string) error {
43 if args.NoYolo {
44 fmt.Fprintln(os.Stderr, "warning: --no-yolo has no effect in rpc mode (no interactive prompt available); tools will run without confirmation")
45 }
46 r, err := Resolve(args, true)
47 if err != nil {
48 return err
49 }
50
51 // Extensions: same lifecycle as interactive mode, minus the
52 // host-hooks integration. Notify/Display calls from extensions
53 // emit RPC events instead of TUI lines so any consumer can react.
54 extHooks := &rpcExtHooks{}
55 extMgr := extensions.New(ZotHome(), r.CWD, version, r.Provider, r.Model, extHooks)
56 for _, e := range extMgr.LoadExplicit(ctx, args.Exts) {
57 fmt.Fprintln(os.Stderr, "extension load:", e)
58 }
59 if !args.NoExt {
60 for _, e := range extMgr.Discover(ctx) {
61 fmt.Fprintln(os.Stderr, "extension load:", e)
62 }
63 }
64 extMgr.WaitForReady(3 * time.Second)
65 defer extMgr.Stop(2 * time.Second)
66 r.MergeExtensionTools(&extToolAdapter{mgr: extMgr})
67
68 ag := r.NewAgent()
69 ag.BeforeToolExecute = func(call provider.ToolCallBlock) (bool, string, json.RawMessage) {
70 r := extMgr.InterceptToolCall(ctx, call.ID, call.Name, call.Arguments)
71 if r.Block {
72 return false, r.Reason, nil
73 }
74 return true, "", r.ModifiedArgs
75 }
76 ag.BeforeTurn = func(step int) (bool, string) {
77 r := extMgr.InterceptTurnStart(ctx, step)
78 return !r.Block, r.Reason
79 }
80 ag.BeforeAssistantMessage = func(text string) (bool, string, string) {
81 r := extMgr.InterceptAssistantMessage(ctx, text)
82 if r.Block {
83 return false, r.Reason, ""
84 }
85 return true, "", r.ReplaceText
86 }
87 ag.OnEvent = func(ev core.AgentEvent) { fanoutAgentEvent(extMgr, ev) }
88
89 // /reload-ext hot-reload callback (also triggered via rpc
90 // `reload_ext` if/when added). Rebuilds the tool registry on the
91 // current agent so freshly-registered extension tools become
92 // callable without restarting the rpc process.
93 adapter := &extToolAdapter{mgr: extMgr}
94 extMgr.SetOnReload(func() {
95 resolved, err := Resolve(args, true)
96 if err != nil {
97 return
98 }
99 resolved.MergeExtensionTools(adapter)
100 ag.SetTools(resolved.ToolRegistry)
101 })
102
103 extMgr.EmitEvent(extproto.EventFromHost{Event: "session_start"})
104
105 server := &rpcServer{
106 ctx: ctx,
107 args: args,
108 agent: ag,
109 provider: r.Provider,
110 model: r.Model,
111 out: os.Stdout,
112 version: version,
113 }
114 extHooks.server = server
115 return server.run(os.Stdin)
116}
117
118// rpcExtHooks implements extensions.HostHooks for the headless RPC
119// loop. Notify and Display surface as `event` frames so any RPC
120// client can render them; Submit and Insert are no-ops because the
121// RPC loop has no editor and the prompt comes from the client.
122type rpcExtHooks struct {
123 server *rpcServer
124}
125
126func (h *rpcExtHooks) Notify(extName, level, message string) {
127 if h.server != nil {
128 h.server.writeEvent(map[string]any{
129 "type": "ext_notify",
130 "extension": extName,
131 "level": level,
132 "message": message,
133 })
134 }
135}
136func (h *rpcExtHooks) Display(extName, text string) {
137 if h.server != nil {
138 h.server.writeEvent(map[string]any{
139 "type": "ext_display",
140 "extension": extName,
141 "text": text,
142 })
143 }
144}
145func (h *rpcExtHooks) Submit(string) {} // ignored in rpc mode
146func (h *rpcExtHooks) Insert(string) {} // ignored in rpc mode
147
148type rpcServer struct {
149 ctx context.Context
150 args Args
151 agent *core.Agent
152 provider string
153 model string
154 out io.Writer
155 version string
156
157 writeMu sync.Mutex
158 turnMu sync.Mutex // serialises one prompt at a time
159 activeCancel context.CancelFunc
160 authed bool
161
162 // inFlight tracks long-running command goroutines so run() can
163 // wait for them before returning when stdin closes. Without this,
164 // piping a single 'prompt' command into 'zot rpc' would race the
165 // process exit against the agent loop and the prompt would never
166 // produce output.
167 inFlight sync.WaitGroup
168}
169
170// run reads NDJSON commands from in and dispatches them. Returns when
171// in is closed AND every in-flight long-running command (prompt /
172// compact) has finished, so a quick `echo cmd | zot rpc` invocation
173// still produces full output before the process exits.
174func (s *rpcServer) run(in io.Reader) error {
175 requireToken := os.Getenv("ZOTCORE_RPC_TOKEN") != ""
176 s.authed = !requireToken
177
178 sc := bufio.NewScanner(in)
179 sc.Buffer(make([]byte, 0, 64*1024), 16*1024*1024)
180 for sc.Scan() {
181 line := strings.TrimSpace(sc.Text())
182 if line == "" {
183 continue
184 }
185 var head struct {
186 Type string `json:"type"`
187 ID string `json:"id,omitempty"`
188 }
189 if err := json.Unmarshal([]byte(line), &head); err != nil {
190 s.writeError("", "", fmt.Sprintf("malformed json: %v", err))
191 continue
192 }
193 if !s.authed {
194 if head.Type != "hello" {
195 s.writeError(head.ID, head.Type, "auth required: send hello with token first")
196 continue
197 }
198 var hello struct {
199 Token string `json:"token"`
200 }
201 _ = json.Unmarshal([]byte(line), &hello)
202 if hello.Token != os.Getenv("ZOTCORE_RPC_TOKEN") {
203 s.writeError(head.ID, head.Type, "invalid token")
204 return fmt.Errorf("rpc: bad auth token")
205 }
206 s.authed = true
207 s.writeResponse(head.ID, head.Type, map[string]any{
208 "protocol_version": 1,
209 "version": s.version,
210 "provider": s.provider,
211 "model": s.model,
212 })
213 continue
214 }
215 s.dispatch(head.Type, head.ID, []byte(line))
216 }
217 err := sc.Err()
218 s.inFlight.Wait()
219 return err
220}
221
222// dispatch routes a command. Long-running commands (prompt, compact)
223// run on their own goroutine so the read loop stays responsive.
224func (s *rpcServer) dispatch(cmd, id string, raw []byte) {
225 switch cmd {
226 case "hello":
227 s.writeResponse(id, cmd, map[string]any{
228 "protocol_version": 1,
229 "version": s.version,
230 "provider": s.provider,
231 "model": s.model,
232 })
233 case "prompt":
234 var req struct {
235 Message string `json:"message"`
236 Images []struct {
237 MimeType string `json:"mime_type"`
238 Data []byte `json:"data"`
239 } `json:"images"`
240 }
241 if err := json.Unmarshal(raw, &req); err != nil {
242 s.writeError(id, cmd, err.Error())
243 return
244 }
245 s.inFlight.Add(1)
246 go func() {
247 defer s.inFlight.Done()
248 s.runPrompt(id, req.Message, req.Images)
249 }()
250
251 case "abort":
252 if c := s.takeCancel(); c != nil {
253 c()
254 }
255 s.writeResponse(id, cmd, nil)
256
257 case "compact":
258 s.inFlight.Add(1)
259 go func() {
260 defer s.inFlight.Done()
261 s.runCompact(id)
262 }()
263
264 case "get_state":
265 s.writeResponse(id, cmd, s.snapshotState())
266
267 case "get_messages":
268 s.writeResponse(id, cmd, map[string]any{
269 "messages": messagesToJSON(s.agent.Messages()),
270 })
271
272 case "clear":
273 s.agent.SetMessages(nil)
274 s.writeResponse(id, cmd, nil)
275
276 case "set_model":
277 var req struct {
278 Model string `json:"model"`
279 }
280 if err := json.Unmarshal(raw, &req); err != nil {
281 s.writeError(id, cmd, err.Error())
282 return
283 }
284 if _, err := provider.FindModel(s.provider, req.Model); err != nil {
285 s.writeError(id, cmd, err.Error())
286 return
287 }
288 s.agent.Model = req.Model
289 s.model = req.Model
290 s.writeResponse(id, cmd, map[string]any{"model": req.Model})
291
292 case "get_models":
293 out := []map[string]any{}
294 for _, m := range provider.ModelsForProvider(s.provider) {
295 out = append(out, map[string]any{
296 "id": m.ID,
297 "provider": m.Provider,
298 "context_window": m.ContextWindow,
299 "max_output": m.MaxOutput,
300 "reasoning": m.Reasoning,
301 })
302 }
303 s.writeResponse(id, cmd, map[string]any{"models": out})
304
305 case "ping":
306 s.writeResponse(id, cmd, map[string]any{"pong": true})
307
308 default:
309 s.writeError(id, cmd, "unknown command")
310 }
311}
312
313// runPrompt executes a single prompt turn and streams events out.
314// Holds turnMu so a second concurrent prompt blocks until this one
315// finishes; the user can abort with the abort command.
316func (s *rpcServer) runPrompt(id, message string, images []struct {
317 MimeType string `json:"mime_type"`
318 Data []byte `json:"data"`
319}) {
320 s.turnMu.Lock()
321 defer s.turnMu.Unlock()
322
323 subCtx, cancel := context.WithCancel(s.ctx)
324 s.setCancel(cancel)
325 defer s.setCancel(nil)
326
327 s.writeResponse(id, "prompt", map[string]any{"started": true})
328
329 imgs := make([]provider.ImageBlock, 0, len(images))
330 for _, im := range images {
331 imgs = append(imgs, provider.ImageBlock{MimeType: im.MimeType, Data: im.Data})
332 }
333
334 err := s.agent.Prompt(subCtx, message, imgs, func(ev core.AgentEvent) {
335 // EvDone is emitted by the agent loop and we re-emit our own
336 // 'done' below; suppressing it here avoids duplicate frames.
337 if _, ok := ev.(core.EvDone); ok {
338 return
339 }
340 s.writeEvent(modes.EventToJSON(ev))
341 })
342 // Don't emit a stand-alone error event for cancellation; the prior
343 // turn_end with stop=aborted already carries that signal.
344 if err != nil && !errors.Is(err, context.Canceled) {
345 s.writeEvent(map[string]any{"type": "error", "message": err.Error()})
346 }
347 s.writeEvent(map[string]any{"type": "done"})
348}
349
350// runCompact mirrors runPrompt but for compaction.
351func (s *rpcServer) runCompact(id string) {
352 s.turnMu.Lock()
353 defer s.turnMu.Unlock()
354
355 subCtx, cancel := context.WithCancel(s.ctx)
356 s.setCancel(cancel)
357 defer s.setCancel(nil)
358
359 s.writeResponse(id, "compact", map[string]any{"started": true})
360 summary, err := s.agent.Compact(subCtx, 4, nil)
361 if err != nil {
362 if !errors.Is(err, context.Canceled) {
363 s.writeEvent(map[string]any{"type": "error", "message": err.Error()})
364 }
365 return
366 }
367 s.writeEvent(map[string]any{
368 "type": "compact_done",
369 "summary": summary,
370 })
371}
372
373// snapshotState builds the get_state response.
374func (s *rpcServer) snapshotState() map[string]any {
375 cum := s.agent.Cost()
376 return map[string]any{
377 "provider": s.provider,
378 "model": s.model,
379 "cwd": s.args.CWD,
380 "message_count": len(s.agent.Messages()),
381 "busy": s.activeCancel != nil,
382 "usage": map[string]any{
383 "input": cum.InputTokens,
384 "output": cum.OutputTokens,
385 "cache_read": cum.CacheReadTokens,
386 "cache_write": cum.CacheWriteTokens,
387 "cost_usd": cum.CostUSD,
388 },
389 }
390}
391
392// ---- write helpers (single-line JSON, mutex-guarded) ----
393
394func (s *rpcServer) writeResponse(id, cmd string, data any) {
395 frame := map[string]any{
396 "type": "response",
397 "command": cmd,
398 "success": true,
399 }
400 if id != "" {
401 frame["id"] = id
402 }
403 if data != nil {
404 frame["data"] = data
405 }
406 s.write(frame)
407}
408
409func (s *rpcServer) writeError(id, cmd, msg string) {
410 frame := map[string]any{
411 "type": "response",
412 "command": cmd,
413 "success": false,
414 "error": msg,
415 }
416 if id != "" {
417 frame["id"] = id
418 }
419 s.write(frame)
420}
421
422func (s *rpcServer) writeEvent(payload map[string]any) {
423 s.write(payload)
424}
425
426func (s *rpcServer) write(v any) {
427 b, err := json.Marshal(v)
428 if err != nil {
429 return
430 }
431 s.writeMu.Lock()
432 defer s.writeMu.Unlock()
433 _, _ = s.out.Write(b)
434 _, _ = s.out.Write([]byte("\n"))
435}
436
437func (s *rpcServer) setCancel(c context.CancelFunc) {
438 s.writeMu.Lock()
439 s.activeCancel = c
440 s.writeMu.Unlock()
441}
442
443func (s *rpcServer) takeCancel() context.CancelFunc {
444 s.writeMu.Lock()
445 defer s.writeMu.Unlock()
446 c := s.activeCancel
447 s.activeCancel = nil
448 return c
449}
450
451// messagesToJSON serialises a transcript using the same schema as the
452// --json event mode for cross-format consistency.
453func messagesToJSON(msgs []provider.Message) []map[string]any {
454 out := make([]map[string]any, 0, len(msgs))
455 for _, m := range msgs {
456 out = append(out, map[string]any{
457 "role": string(m.Role),
458 "content": modes.ContentToJSON(m.Content),
459 "time": m.Time,
460 })
461 }
462 return out
463}